docs: simplify project header and add DeepWiki

.devin/wiki.json

@@ -0,0 +1,33 @@
+{
+  "repo_notes": [
+    {
+      "content": "Use current source and README.md, docs/EXPERIMENTS.md and each result record as authority. StateEval is an independent evaluation harness, not the transaction service or a model runtime. Acting through agent/business endpoints and judging through a separate read-only SQL account are different paths; /api/eval/state and audit endpoints are diagnostic, not the oracle. Core contains no CityBuddy imports or SQL; adapters own business knowledge. Preserve hard-gate order: final business state, forbidden side effects, permission violations. Treat tool trajectories as diagnostics. Keep the historical support-agent campaign and the current ShopMate buyer calibration separate, each with its original code versions, model configuration, task set and denominator. Historical 55/300 versus 0/300 does not describe current ShopMate; current foreign-order pairs were 0/3 versus 0/3 and did not reach a sensitive write, so they do not demonstrate incremental ownership-check effectiveness. Two own-order positive controls passed; repeated confirmation is not another model trial. A REQUESTED refund is an accepted request, not funds arrival. Never imply model aliases are immutable snapshots, ablations affect production, or local pilot results prove general safety."
+    }
+  ],
+  "pages": [
+    {
+      "title": "Overview and evaluation design",
+      "purpose": "Explain the authorization question, independent acting/judging paths, gate order and relationship to external CityBuddy and ShopMate."
+    },
+    {
+      "title": "Local setup and reproducible runs",
+      "purpose": "Document the current Makefile, pinned runtime, sibling checkouts, positive controls, isolated fixtures, source-clean measured revisions and safe cleanup from active guides/scripts."
+    },
+    {
+      "title": "Core engine and target adapters",
+      "purpose": "Describe core task/turn/assertion/gate types, adapter seam, trial execution, CityBuddy and ShopMate adapters, real buyer factories and independent SQL grading."
+    },
+    {
+      "title": "Campaigns and result interpretation",
+      "purpose": "Explain ownership/session experiments, equal-budget controls and denominators. Keep historical formal results, current buyer calibration and positive controls separate; include non-activated and inconclusive outcomes."
+    },
+    {
+      "title": "Artifacts, SQL evidence and failure handling",
+      "purpose": "Describe raw transcripts, independent before/after SQL, outcome files, retained-fixture flags and operational failures without treating tool logs or diagnostic APIs as business truth."
+    },
+    {
+      "title": "Testing, methodology and contribution",
+      "purpose": "Cover boundary checks, unit/integration tests, Actions pinned dependencies, contribution rules and prior-art references. Explain limits and reproduction without fabricating coverage or results."
+    }
+  ]
+}

README.md

@@ -1,10 +1,9 @@
-![StateEval · Evaluate agent actions through independent business-state checks](docs/assets/cover.png)
-
 # StateEval
 
 **English** · [简体中文](README.zh-CN.md) · [Contributing](CONTRIBUTING.md)
 
 [![check](https://github.com/ChanTso/state-eval/actions/workflows/check.yml/badge.svg?branch=main)](https://github.com/ChanTso/state-eval/actions/workflows/check.yml)
+[![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/ChanTso/state-eval)
 
 **Let the agent act. Use independent SQL to check what it actually changed.**
 

README.zh-CN.md

@@ -1,10 +1,9 @@
-![StateEval · 从独立业务终态判断 Agent 行为](docs/assets/cover.png)
-
 # StateEval
 
 [English](README.md) · **简体中文** · [贡献指南](CONTRIBUTING.md)
 
 [![check](https://github.com/ChanTso/state-eval/actions/workflows/check.yml/badge.svg?branch=main)](https://github.com/ChanTso/state-eval/actions/workflows/check.yml)
+[![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/ChanTso/state-eval)
 
 **让 Agent 执行业务，用独立 SQL 检查它最终做了什么。**